The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to analyze external data such as AI plans, RAG results, and evidence sets. This processing of untrusted data represents a potential attack surface for indirect prompt injection.
Ingestion points: The skill processes user-supplied LLM plans, results, and evidence sets as specified in SKILL.md.
Boundary markers: The instructions in prompts/ai-safety-testing.md require the agent to strictly separate confirmed facts, inferences, and human decisions, creating logical boundaries.
Capability inventory: The skill generates text-based reports (Markdown, CSV, JSON). There are no instructions for the agent to execute code or perform network operations based on the analyzed data.
Sanitization: The skill includes explicit instructions to use least privilege, masked data, mocks, and isolation when handling sensitive user or production data.
[COMMAND_EXECUTION]: The skill package includes local utility scripts used for automated evaluation and verification.
Evidence: evals/fixtures/check-domain-decisions.sh and evals/fixtures/check_domain_decisions.py are provided to validate that the agent's output matches expected safety decision formats. These are standard testing components and do not represent a runtime risk for the end-user.

ai-safety-testing